AI Brand Voice: Build a Guide From Your Actual Writing

An AI brand voice guide is a set of writing decisions supported by examples. It tells the assistant what to do, shows what that looks like, and gives the person reviewing the draft something specific to check.

You can start with writing you already have. Pick a piece that sounds like you. Point to the sentence that makes it work. Explain the choice behind that sentence. Then test whether the same choice helps with a new task.

That is more useful than handing over a list of personality adjectives and hoping the assistant interprets them the way you do.

At Definitely Human, our voice reference separates private conversations, communication with people who know us, and writing for strangers. Each situation has a different opening. The rules come from examples of work that was actually sent, published or delivered.

This guide explains that method. The demonstration passages below are deliberately constructed examples, not reported model outputs or customer results.

What should an AI brand voice guide contain?

Start with four things: the reader, the writing choice, an example, and an explanation of why the example fits.

“Direct” is a useful description, but it leaves decisions open. Does direct mean a short opening? A strong opinion? No introduction? In a private message, it might mean answering the question before explaining the background. In a guide, it might mean stating what the reader will be able to do.

Turn the description into a decision someone could apply.

These are example rules, not a universal brand personality. A technical journal and a founder’s newsletter can make different choices and both be good.

Your AI brand voice guide should make your choices visible. It should also distinguish style from truth. A warm tone cannot authorize a new guarantee. A confident sentence cannot turn an assumption about a customer into a verified fact.

Keep facts about the company and its offer in a separate AI context document. The voice reference explains how to express those facts.

Which writing samples should you use?

Choose material that represents the work you want the assistant to help produce. A polished article is evidence for an article. It may be a poor model for a short reply to a client.

Look for sent emails, published posts, approved pages and transcripts of delivered talks. For each sample, record the situation: who it addressed, what the reader already knew, and what the writer wanted to happen next.

Also record what you like about it. You may want the clear explanation without the joke. You may like the opening but no longer agree with the offer. Supplying a sample without those boundaries can invite the assistant to repeat the wrong part.

We encountered that problem in our own documentation. A prepared workshop script looked like a strong voice sample. The delivered transcript differed from the script in an important opening choice. Treating the script as evidence would have recorded something Martin planned to say as something he actually said.

The correction was simple: when the question is how someone speaks, the delivered transcript takes precedence over the unused script. The script can still help explain structure. It cannot establish delivery.

That distinction applies to written work too. An abandoned draft is not evidence of an approved voice. A document full of tracked changes may contain both the choices you kept and the ones you rejected.

Give each source a status before extracting patterns:

• Used: sent, published or delivered, and still representative.
• Proposed: a direction you want to test but have not adopted.
• Rejected: a useful counterexample, with a reason attached.
• Outdated: once valid, now superseded by a named decision.

Keep only material you are entitled to use, and remove private details that do not help explain the writing. You can study the structure of a client reply without carrying the client’s account details into every future task.

How do you turn examples into voice rules?

Read a sample with one question: what choice could another writer repeat without copying this sentence?

A phrase can be distinctive, but repeating it everywhere quickly becomes imitation. Look at the underlying decision: where the explanation starts, how uncertainty is expressed, how an example is introduced, or how the writer asks for a response.

Our published line “Build marketing that learns” is short and points toward a specific idea. It does not mean every explanation should be four words. A slogan has a different job from a tutorial.

Likewise, our preference for short paragraphs does not mean a sentence should be broken in half merely to create white space. The useful rule is to keep the paragraph focused and let the reader see the progression.

Use a small evidence table when building an AI brand voice reference:

For each rule, add a counterexample. Explain the failure precisely. “This is bad” teaches little. “This opening assumes the reader attended our workshop” identifies the problem and the correction.

Do not turn every preference into an absolute. If your strongest samples contain both short and long sentences, a universal short-sentence rule would erase something useful. Write down the condition under which the choice works.

How should the voice change for different readers?

One brand can sound recognizable without using the same opening everywhere.

Our own reference distinguishes three relationships. In a private conversation, the writer can refer to something that person said. With an opted-in audience, shared context may support a more personal opening. For a stranger arriving from search, the writing has to establish its usefulness without assuming a relationship.

Those distinctions are more actionable than “email voice” and “blog voice.” An email sent individually to strangers is still an introduction. A long note to an existing client can assume knowledge that a public article cannot.

Before asking for a draft, supply a sentence describing the relationship:

This reader found the guide through search. They know what AI writing is, but they do not know our framework or our services.

That instruction gives the writer a reason to define unfamiliar concepts and skip private references.

For a different task, the instruction might be:

This is a reply to someone who has already reviewed the proposal. Answer their question about handover before introducing anything new.

The underlying voice may still be clear and specific. The opening and amount of explanation change because the reader’s needs change.

Do not mistake context changes for inconsistency. An apology, an educational guide and a sales proposal should not carry identical emotional weight. An AI brand voice guide should preserve judgment about the situation.

How do you test whether the guide works?

Choose one task with a clear audience, a defined subject and a factual source. Keep that task stable while you compare drafts. If you change the offer, audience and instructions between attempts, you will not know what caused the difference.

Review against named rules before asking whether the result “feels right.” Taste still matters, but a specific diagnosis makes the next revision easier.

Here is an illustrative editing exercise, constructed for this guide. It is not a measured before-and-after AI experiment.

Task: introduce a guide that helps a consulting founder turn customer questions into content ideas.

Passage to review:

In today’s rapidly evolving landscape, innovative consultants must leverage the transformative power of artificial intelligence to unlock their content potential.

Editorial correction:

Start with a question a customer has already asked. Use it to choose one useful explanation you can support with evidence.

The second passage names an action, identifies the input and sets a boundary. The first could introduce almost any AI article. That is the reason to change it. Counting banned words alone would miss the larger problem.

Now ask what the corrected passage still lacks. It does not explain how to choose among several questions. It does not show the evidence. Those are content gaps, not voice defects, and the rest of the guide must resolve them.

For an actual test, save the task, the guide version, the output and your corrections together. Keep the generated draft separate from your edited version. That lets you distinguish what the system produced from what the editor supplied.

Try another task after the first correction. A rule that improves one opening might make every subsequent opening repetitive. You are testing whether the rule travels, not just whether you can repair a single paragraph.

What should you change when the draft still sounds wrong?

Identify the smallest useful correction. Did the assistant miss a rule? Was the example unsuitable? Did the instruction conflict with another document? Or is the rule itself too broad?

Those problems require different fixes. Repeating an instruction more forcefully will not resolve a contradiction between two current-looking references.

Record a correction only when you understand its scope. If you cut a sentence because this particular reader knows the background, do not create a global rule to remove all background explanations.

This is where an AI brand voice guide becomes maintained working knowledge. Keep a short change note: what happened, what rule changed, and which output you used to check it. The note need not become another administrative project. It just needs to preserve the reason behind the change.

Avoid accumulating a wall of prohibitions. If ten instructions describe the same problem, rewrite them into a clearer rule and keep the best examples. A reference that nobody can inspect is difficult to maintain.

How do you keep the guide useful across tools?

Keep one editable master reference in a format your team can read and move. Plain text or Markdown works well for many workflows, but the format matters less than having a clear owner and a known current version.

Separate the reference from the task. The reference holds recurring writing choices. The task supplies today’s audience, subject, format and source material.

For example, a request to draft a guide should identify which voice reference to use and include the business facts needed for that guide. It should not depend on a previous conversation being remembered.

Tool behavior varies. Some setups can reuse project materials; others require you to supply them again. An uploaded document is not proof that the next task will use the right passage. Test the actual setup with a fresh task and inspect the result.

Also distinguish supplying context from training model weights. Most teams beginning this process are providing instructions and examples to an existing model. They are not building a new model, and the guide should not imply otherwise.

When you change tools, run the same small test again. Do not assume identical output from identical instructions. Your advantage is that the choices and evidence remain available to inspect.

What should you build first?

Start with a working AI brand voice reference for one recurring task. Choose a task that you can review well, such as a short educational post or a response to a familiar customer question.

Collect representative material, extract a few supported rules, and write down what should vary with the reader. Then review an output and retain the correction that matters.

You have a usable first version when another person can explain why a passage fits or fails without appealing only to instinct. That does not mean the writing is finished forever. It means the decisions can be discussed and improved.

Once the voice reference is usable, connect it to the AI marketing workflow. That guide shows where source selection, factual review and approval belong around the writing step.

If you need a subject to practice on, the Content Idea Engine starts with customer questions. Keep the test small enough that you can see which correction improved the work.